Detection of Linear Contrails with a Morphological Algorithm and with Deep Neural Networks

The climate impact of aviation can be separated into CO2 and non-CO2 effects, with the latter being potentially larger than the former. In this context we are more specifically interested in condensation trails (hereafter contrails) and induced cirrus. Monitoring contrail formation and evolution is necessary to understand their radiative effects and help the aviation industry to transition towards a more sustainable activity. Current research aimed at detecting contrails is mostly based on geostationary satellite images because they allow to follow the contrail over a long period of time. However a major shortcoming of this approach, due to the current spatial resolution of geostationary imagers, is that the contrail formation phase cannot be detected and larger, but older, contrails cannot always be attributed to the flights that produced them. To circumvent this problem and observe the contrail formation phase, we use a ground-based hemispheric camera with a two-minute sampling rate as a complementary source of information. As a first step, we have developed a traditional morphological algorithm that will help preparing a sufficiently large labelled database as required to train a deep-learning algorithm. This algorithm aims to detect whether each aircraft that passes in the field of view of the camera (as monitored from an ADSB radar) produces a contrail or not and, whenever possible, track the contrail across successive images. We are thus able to relate contrail formation and evolution with aircraft type, flight altitude and weather conditions.  We consider all weather conditions except completely cloudy conditions that prevent contrails from being observed. The performance of this algorithm is evaluated against a database that was manually annotated consisting of 400 images with 407 contrails. We find a specificity of 97\%, i.e. there are few false detections, but a sensitivity of about 55\%, i.e. it is missing a significant fraction of contrail that were annotated manually. An analysis of several years of contrail detection will be presented to determine precisely the fraction of contrail-producing flights and the  weather conditions associated with short-lived (less than 2 min) and longer-lived(more than 2 min) contrails. Additionally to this approach, which misses part of the young contrails and does not detect contrails formed outside the field of view of the camera, we have trained deep neural networks such as Unet and DeeplabV3, on a database of 1600 images in order to overcome those limits. Our preliminary results show a good performance on young contrails, with an improved detection capability, in particular for contrails formed outside of the camera field of view. The deep neural networks also work better for old contrails but may confuse very old contrails with background cloud features.